• The Journal of the Acoustical Society of Korea
• /
• v.35 no.6
• /
• pp.501-509
• /
• 2016

We propose a method to improve the robustness of speaker verification on short test utterances. The accuracy of the state-of-the-art i-vector/probabilistic linear discriminant analysis systems can be degraded when testing utterance durations are short. The proposed method compensates for utterance variations of short test feature vectors using deep neural networks. We design three different types of DNN (Deep Neural Network) structures which are trained with different target output vectors. Each DNN is trained to minimize the discrepancy between the feed-forwarded output of a given short utterance feature and its original long utterance feature. We use short 2-10 s condition of the NIST (National Institute of Standards Technology, U.S.) 2008 SRE (Speaker Recognition Evaluation) corpus to evaluate the method. The experimental results show that the proposed method reduces the minimum detection cost relative to the baseline system.

• Phonetics and Speech Sciences
• /
• v.9 no.2
• /
• pp.85-94
• /
• 2017

Conventional speaker verification systems using time delay neural network, identity vector and probabilistic linear discriminant analysis (TDNN-Ivector-PLDA) are known to be very effective for verifying long-duration speech utterances. However, when test utterances are of short duration, duration mismatch between enrollment and test utterances significantly degrades the performance of TDNN-Ivector-PLDA systems. To compensate for the I-vector mismatch between long and short utterances, this paper proposes to use probabilistic linear discriminant analysis (PLDA) model adaptation with augmented data. A PLDA model is trained on vast amount of speech data, most of which have long duration. Then, the PLDA model is adapted with the I-vectors obtained from short-utterance data which are augmented by using vocal tract length perturbation (VTLP). In computer experiments using the NIST SRE 2008 database, the proposed method is shown to achieve significantly better performance than the conventional TDNN-Ivector-PLDA systems when there exists duration mismatch between enrollment and test utterances.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.1
• /
• pp.42-47
• /
• 2003

This paper proposes the method to generate new test data using the sample shift of the start frame for SPRT(sequential probability ratio test) in speaker verification. The SPRT method is a effective algorithm that can reduce the test computational complexity. However, in making the decision procedure, SPRT can be executed on the assumption that the input samples are usually to be i.i.d. (Independent and Identically Distributed) samples from a probability density function (pdf), also it's not suitable method to apply for the short utterance. The proposed method can achieve SPRT regardless of the utterance length of the test data because it is method to generate the new test data through the sample shift of start frame. Also, the correlation property of data to be considered in the SPRT method can be effectively removed by employing the principal component analysis. Experimental results show that the proposed method increased the computational complexity of data for sample shift a little, but it has a good performance result more than a conventional method above the average 0.7% in EER (equal error rate).

• Proceedings of the IEEK Conference
• /
• 2009.05a
• /
• pp.37-39
• /
• 2009

In the research of speech recognition, locating the beginning and end of a speech utterance in a background of noise is of great importance. Since the background noise presenting to record will introduce disturbance while we just want to get the stationary parameters to represent the corresponding speech section, in particular, a major source of error in automatic recognition system of isolated words is the inaccurate detection of beginning and ending boundaries of test and reference templates, thus we must find potent method to remove the unnecessary regions of a speech signal. The conventional methods for speech endpoint detection are based on two simple time-domain measurements - short-time energy, and short-time zero-crossing rate, which couldn't guarantee the precise results if in the low signal-to-noise ratio environments. This paper proposes a novel approach that finds the Lyapunov exponent of time-domain waveform. This proposed method has no use for obtaining the frequency-domain parameters for endpoint detection process, e.g. Mel-Scale Features, which have been introduced in other paper. Comparing with the conventional methods based on short-time energy and short-time zero-crossing rate, the novel approach based on time-domain Lyapunov Exponents(LEs) is low complexity and suitable for Digital Isolated Word Recognition System.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.1
• /
• pp.64-68
• /
• 2020

In a conventional speech recognition system using Gaussian Mixture Model-Hidden Markov Model (GMM-HMM), the cepstral feature normalization method based on pole filtering was effective in improving the performance of recognition of short utterances in noisy environments. In this paper, the usefulness of this method for the state-of-the-art speech recognition system using Deep Neural Network (DNN) is examined. Experimental results on AURORA 2 DB show that the cepstral mean and variance normalization based on pole filtering improves the recognition performance of very short utterances compared to that without pole filtering, especially when there is a large mismatch between the training and test conditions.

• Journal of the Korean Institute of Telematics and Electronics C
• /
• v.34C no.5
• /
• pp.55-65
• /
• 1997

In many cases, a dialogue system can't extract the correct analysis information of a user's spoken utterance, because of its own ungrammatical components. Therefore, in order to perform a correct before it performs the syntactic processing. In this paper, we use a real dialogue corpus and classify these ungrammatical errors as 4 categories : phonological, syntactic, semantic errors that consist of speech reparis and inversions, and propose an algorithm to detect and correct the errors. In short, this paper proposes a method to detect and correct the speech repairs and inversions that are classified as the phonological and syntactic errors to implement a robust dialogue system. And, through the test of real dialogue data, this paper shows an efficiency of the proposed algorithm.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.46 no.5
• /
• pp.8-14
• /
• 2009

In the research field of speech recognition, pinpointing the endpoints of speech utterance even with the presence of background noise is of great importance. These noise present during recording introduce disturbances which complicates matters since what we just want is to get the stationary parameters corresponding to each speech section. One major cause of error in automatic recognition of isolated words is the inaccurate detection of the beginning and end boundaries of the test and reference templates, thus the necessity to find an effective method in removing the unnecessary regions of a speech signal. The conventional methods for speech endpoint detection are based on two linear time-domain measurements: the short-time energy, and short-time zero-crossing rate. They perform well for clean speech but their precision is not guaranteed if there is noise present, since the high energy and zero-crossing rate of the noise is mistaken as a part of the speech uttered. This paper proposes a novel approach in finding an apparent threshold between noise and speech based on Lyapunov Exponents (LEs). This proposed method adopts the nonlinear features to analyze the chaos characteristics of the speech signal instead of depending on the unreliable factor-energy. The excellent performance of this approach compared with the conventional methods lies in the fact that it detects the endpoints as a nonlinearity of speech signal, which we believe is an important characteristic and has been neglected by the conventional methods. The proposed method extracts the features based only on the time-domain waveform of the speech signal illustrating its low complexity. Simulations done showed the effective performance of the Proposed method in a noisy environment with an average recognition rate of up 92.85% for unspecified person.